Ask three people what AI for WhatsApp means and you get three different answers. One is thinking of the AI app they message from their own phone. Another means the automatic "thanks, we'll get back to you" that fires from a shop's business number. The third means a WhatsApp AI chat that answers a customer's real question — your price, your next free slot, whether you cover their block. This guide is about the third: putting real WhatsApp artificial intelligence on the number your customers already message.
It is a practical piece rather than a product tour: what you need, in what order, what to feed the assistant, and where to draw the line between it and you. If you run a small business in Singapore and most enquiries arrive on WhatsApp, this is the version worth doing properly.
Three different things people call AI for WhatsApp
The first is consumer AI that you message yourself. Some AI companies run an account you can save as a contact and message like a friend, which is what most people mean by ChatGPT on WhatsApp. It is useful for drafting or translating a message, but it is a separate consumer service on your personal phone: not attached to your business number, never shown your price list, and not something your customers are talking to.
The second is auto-replies on a business number: a greeting message, an away message outside your hours, maybe a short menu of options. That is automation, but it is fixed text, and it fires the same words at every question.
The third is an actual assistant on your business number, and it is a different category of thing. If you are still weighing up which kind of bot you need at all, our guide to choosing a WhatsApp chatbot compares them side by side.
- It reads the question the customer actually asked instead of matching a keyword.
- It answers from your prices, policies and documents rather than a generic script.
- It can look at your calendar and hold a real appointment slot.
- It knows when to stop and pass the conversation to a human.
Why the WhatsApp Business app's auto-replies fall short
The free WhatsApp Business app is a sensible first step, but canned replies buy time rather than answer the question. A customer asks how much a full colour costs and whether you have Saturday morning, and receives "Thanks for your message, we will get back to you during business hours." She has already messaged two other salons, and one of them answered properly. Speed only counts if the reply contains the answer, as we argued in reply to WhatsApp leads in 30 seconds.
There is a second, quieter problem: the app lives on one phone. Quick replies are still you typing, and nothing happens at eleven at night unless you are awake. That is the ceiling of ordinary WhatsApp automation. It can acknowledge a message, but it cannot resolve one.
What you need to connect AI to a business number
Three things, none of them exotic.
- A number on the WhatsApp Business Platform, the official business channel rather than the consumer app. Your existing number can usually be moved onto it, so customers keep messaging the digits they already have.
- A provider that connects to that number and runs the assistant, receiving each message, working out the reply, and sending it back within seconds.
- Your own content: prices, service list, policies, opening hours, and the answers you find yourself typing over and over.
The third is the one owners underestimate. The provider supplies the machinery; your content is what makes the answer correct. An assistant with no price list will be fluent, friendly and wrong.
Step by step: putting a WhatsApp AI chat on your number
The order matters more than the speed.
- Decide which number answers: move your main business line across, or start a second number and redirect to it once you trust the setup.
- Register it on the WhatsApp Business Platform through your provider and complete Meta's business verification, which takes days rather than minutes. Start this early.
- Connect your calendar if you want the assistant to book, so it works from your genuine free time instead of a guess.
- Load your content, then read the first fifty replies it drafts and correct every one that is off. This step decides whether it works.
- Set the handover rule, so a conversation reaches you when it needs a person.
- Go live on real enquiries, starting with the hours you cannot answer anyway: evenings, Sundays and public holidays.
One caution on step one. A number on the Business Platform has historically not been usable in the ordinary app at the same time. That is loosening, but confirm how your provider handles it before moving your main line.
What to feed it so the answers are actually right
An assistant is only as accurate as the material behind it, and most owners already have that material somewhere: a price list, a services PDF, the terms you send new customers. The job is gathering it, not writing it.
- Your current prices, including the conditions that change them, such as length, size, urgency or add-ons.
- What you do not do, which is the fastest way to stop a confident wrong answer.
- Your policies on deposits, cancellation windows, rescheduling and callout charges.
- The areas, estates or branches you cover, and the ones you turn down.
- The five questions you answer most often, written the way you would actually answer them.
Keep one source of truth and edit it there. Raise a price in three documents and forget the fourth, and the assistant will quote the fourth. This matters more the more moving parts you have: a clinic running several practitioners off one calendar needs the assistant to know who does what, not just when a slot is empty.
Set the handover rule, and be honest about the limits
Before you go live, decide what the assistant is not allowed to do. Price negotiation belongs on that list, along with complaints, refunds, anything medical or legal, and any sentence that counts as a promise. An assistant that quotes your published price is helpful. One that agrees to a discount because a customer pushed twice has made a decision that was yours.
Handover should be immediate: the assistant stops, you are notified, and you continue in the same thread without the customer repeating themselves. An AI receptionist that improvises past the edge of what it knows costs you more than one that stops early and asks.
Be realistic about the first few weeks. It will get edge cases wrong, especially unusual requests phrased in an unusual way, and it will sometimes answer a question that was really two. Some tools let you run it silently at first, watching how you reply and learning your wording before it ever speaks to a customer. Either way, budget time for correction rather than expecting it to arrive finished.
An assistant that says "let me check with the owner" is not failing. It is doing the one thing a bad one never does.
Languages, and keeping it PDPA-safe
In Singapore the same shop gets enquiries in English, Mandarin, Malay and Tamil, often shortened, often mixed. A good assistant replies in whatever language the customer wrote in, without being told to switch, and copes with the shorthand people actually use rather than textbook phrasing. If half your customers write in Mandarin, test that before you trust it in public.
The moment an assistant reads customer messages you are processing personal data, and the PDPA applies. Ask any provider three plain questions: are identifiers like NRIC and card numbers stripped before anything is stored or sent to a model, where does the processing physically happen, and are your conversations used to train someone else's model. Our PDPA checklist for AI customer messaging covers the rest. A vendor who cannot answer those clearly has answered them.
How to tell it is working
Message volume is a vanity number. Watch three things instead: how long a first reply takes at nine on a Sunday evening, how many enquiries reach a booking without you touching the thread, and how often you had to correct a reply. The third should fall week after week. If it is not dropping by the second month, the problem is your content, not the technology.
Read the transcripts for the first month even when nothing is wrong. It is the fastest education you will get in what customers really ask.
microcrew is one way to do this. It answers WhatsApp and Telegram enquiries in about 30 seconds, in the customer's language, from your own prices and documents in Google Drive. It books into Google Calendar across several staff and branches, reads the voice notes and photos customers send, and hands over the moment something needs your judgement. NRIC and card numbers are redacted before anything is stored, and the AI processing runs in Singapore under a no-training agreement. Plans are SGD $99, $199 and $399 a month, and access is invite-only for now, so the way in is to request early access.